Abdominal Multi-organ Localization with Adaptive Random Forest in CT Images

Abstract

Medical image localization plays an important role in digital medical research, therapy planning, and delivery. However, the presence of noise and low contrast renders automatic abdominal multi-organ localization an extremely challenging task. In this study, we focus on an adaptive weighted random forest method for abdominal organ localization. Different from the traditional random forest, the proposed method first trains the weighted random regression forest, then iterates multiple weighted random forests to form a stronger one with the Adaptive Boosting technique, and finally performs the final regression through the random forest regressor to get the final bounding box. Our adaptive random forest algorithm can efficiently realize the direct mapping from voxel to organ position and size. Through the quantitative verification of CT scanning data, our method has higher accuracy than the original random forest method and AdaBoost method.